🛡️ Universal Skill Auditor

中文文档向下滚动 | Scroll down for Chinese version

The ultimate automated gatekeeper for your AI Agent ecosystem. As LLMs and AI agents become deeply integrated into our daily workflows, executing third-party, prompt-based "skills" poses a silent but critical security risk. skill-auditor is a universal, platform-agnostic security tool designed to prevent malicious skills from overriding your system instructions, exfiltrating local data, or executing dangerous commands.

✨ Core Feature: Zero-Click Automatic Auditing

Top-tier security should be invisible. skill-auditor features a fully automated interception mechanism. Whenever a new skill or prompt-based program is installed or loaded into your workspace, skill-auditor instantly wakes up in the background. It scans the incoming logic and blocks threats before they can execute—requiring absolutely no manual triggers or prompts from the user.

🔍 The 6-Dimensional Security Check

Our auditor meticulously dissects any incoming skill across six critical vectors:

1. Prompt Injection: Detects attempts to overwrite core system instructions, jailbreak the AI, or alter foundational rules.
2. File System Threats: Flags unauthorized attempts to delete files or write to sensitive system paths.
3. Network Exfiltration: Blocks commands designed to secretly send your local data or environmental variables to external servers.
4. Privilege Escalation: Alerts you if a skill attempts to request sudo or administrator-level access.
5. Social Engineering: Identifies misalignments between what a skill claims to do and its actual hidden behavior.
6. Stealth Behaviors: Uncovers hidden instructions, encoded payloads (e.g., Base64), or malicious commands masked by excessive blank lines.

🚀 Usage

• Automatic Mode (Recommended): Do nothing! Just install or load your skills as usual. skill-auditor acts as a silent sentinel in the background.
• Manual Mode: You can explicitly trigger an audit by commanding your AI: "Audit this skill" or "Run a security check on the newly downloaded tool".

---

🇨🇳 中文文档

为你所有的 AI Agent 和大模型技能库打造的通用“门卫”。

当我们把 AI 深度接入各类开发和业务工作流时，运行来源不明的第三方 Skill（提示词程序）会带来极大的安全隐患。skill-auditor 是一款打破平台壁垒的通用安全审计工具，专为拦截恶意指令、防止越权访问和保护本地数据而生。

✨ 核心亮点：全自动、无感知的安全审查

最好的安全工具是让用户感受不到它的存在。skill-auditor 拥有全自动的静默拦截机制。 无论是哪个平台的 Agent 框架，只要检测到有新的 skill 被安装或写入，审查程序就会在后台瞬间启动。它会在恶意代码生效前完成扫描并阻断威胁，全程无需你开口下达任何指令。

🔍 6 大核心审查维度

像“X光”一样透视每一个外来 Skill：

1. 提示词注入 (Prompt Injection)： 严查是否试图覆盖核心系统指令、越狱 (Jailbreak) 或篡改底层规则。
2. 文件系统高危操作： 拦截越权删除文件或向敏感系统路径写入的动作。
3. 网络数据外泄： 防止本地代码、数据或环境变量被暗中打包发送到外部服务器。
4. 权限提升： 警惕任何试图要求 sudo 或管理员权限的隐藏指令。
5. 社会工程学欺骗： 核对 Skill 的“自我描述”与其“实际执行逻辑”是否一致，防止挂羊头卖狗肉。
6. 隐蔽恶意行为： 揪出藏在大量空行后、或者经过 Base64 等编码伪装的恶意 Payload。

🚀 如何使用

• 全自动模式（推荐）： 保持默认即可。当你导入或安装新 Skill 时，它会自动站岗放哨。
• 手动模式： 随时向你的 AI 助手下达指令：“审查这个 skill” 或 “帮我安全扫描一下刚下载的技能”。